Qwen 3.8 27B might be the most interesting local AI release I've tested this month.

27.78B parameters. 262K context. Apache 2.0. And it beat Claude Opus 4.6 Max on 15/19 launch benchmarks. But I wanted to see what happens outside benchmarks.

Julian Goldie SEOAug 19

“Not caring, it just does stuff.”

Pretty remarkable view of what qwen 3.8 uncensored is doing

Herb ScribnerAug 19

Dflash 2 on M2 Max gave me around 22 tok/s for Qwen 3.8 27B 4Bit from mlx-community

Is this the max I should expect on a M2 Max or the setup required fine tuning? running in oMLX anyone running on M2 Max have similiar results or maybe a bit higher? how good it went for you

Chris WAug 19

You can now build 3D animated websites using Qwen 3.8 27b

Much like the 'build $10k websites using Claude Code', all you need is a solid process and you can achieve similar results Try this out: - Launch Qwen inside - Wire up the higgsfield_ai MCP -

bartslodyczkaAug 19